Page 11 - Read Online

Dababneh et al. Art Int Surg 2024;4:214-32 https://dx.doi.org/10.20517/ais.2024.50 Page 218

[27]
also used a DL network to detect wrist fractures, aiming to assist physicians in their diagnosis . Overall,
there was a 47.0% reduction in misinterpretation rate. These findings are consistent with Cohen et al.’s
study, which noted an increase in wrist fracture detection by non-specialized radiologists when assisted by
[28]
AI . While these studies show promise, it is important to note that the sensitivity of AI is not perfect and
that their performance might be reduced with more complex fractures. In 2022, Hardalaç et al. combined
five ensemble models to create the “wrist fracture detection-combo (WFD-C)”, which had the highest
detection rate compared to other models, with an 86.39% average precision .
[29]

Lysdahlgaard explored a different approach for AI automated fracture detection by investigating the
potential of heat maps . Using the MURA dataset, 20 ML models were used to interpret heat maps
[30]
generated from X-rays. The overall accuracy for all the models combined was 81% for wrist radiographs.
Alammar et al. also used the MURA dataset to collect radiographs of the humerus and wrist to enhance the
performance of a pre-trained convolutional neural network (CNN) model, originally adapted from
ImageNet . The proposed algorithm achieved an AUC of 0.856, thereby outperforming ImageNet models
[31]
that did not receive additional training. Similarly, in 2024, Jacques et al. compared the performance of 23
[32]
radiologists with varying levels of experience with and without AI assistance . BoneView (Gleamer), the
DL model used in this study, achieved an AUC of 0.764 for fracture detection. AI assistance enhanced
radiologists’ sensitivity by 4.5% but did not affect their specificity.

Distal radius fracture
Kim and MacKinnon were among the firsts to apply a CNN specifically to the detection of distal radius
fracture. In their study, a pretrained Inception v3 network was enhanced using a set of lateral wrist
radiographs . While this study successfully demonstrated a proof of concept where the model achieved an
[33]
AUC of 0.954, researchers acknowledged that incorporating a second imaging view could potentially
improve the model’s diagnostic performance. In 2019, Gan et al. further explored the application of AI in
distal radius fracture detection by using anterior-posterior (AP) views instead of lateral projection . They
[34]
trained the Inception-v4 model as a diagnostic tool and the Faster region-based convolutional neural
network (R-CNN) model as an auxiliary algorithm tasked with identifying regions of interest within the
radiographs. The Inception-v4 model achieved an AUC of 0.96 and an overall diagnostic accuracy of 93%.
Their findings also suggest that AI achieves detection rates comparable to those of an experienced
orthopedic surgeon when using AP radiographs. Similarly, Thian et al. utilized the Inception-ResNet and
the Faster R-CNN models, training them on 7356 postero-anterior (PA) and lateral wrist radiographs to
[35]
detect fractures . The combined model achieved a diagnostic accuracy of 91.2% for radius fractures and
96.3% for ulna fractures on the PA and lateral projections, respectively. Oka et al. trained a different DL
model, VGG16, on 498 AP images and 485 lateral radiographs of the distal radius, as well as 491 images of
the styloid process of the ulna. The model demonstrated a diagnostic accuracy of 98.0% (AUC 0.991) for
[36]
distal radius fractures and 91.1% (AUC 0.991) for fractures of the styloid process .

Russe et al. also conducted a study evaluating various AI models for detecting distal radius fractures by
using three models: a classification model to recognize fractures images, a segmentation model to locate
precise fracture boundaries within images, and a detection model to identify fractures . This model
[37]
achieved high accuracies, up to 97%, and effective fracture localization. Likewise, Zhang et al. developed and
[38]
evaluated a DL algorithm for diagnosing distal radius fractures based on X-ray images . Their study
included a total of 3,276 wrist X-ray films. The DL model achieved a high accuracy of 97.03%, with a
sensitivity of 95.70% and a specificity of 98.37%, outperforming both orthopedic and radiology attending
physicians. Anttila et al. employed a segmentation-based U-net model, which accurately identified distal
radius fracture with an AUC of 0.97 for radiographs without casts . Accuracy was better in PA views
[39]